GPQA

Graduate-level Google-Proof Q&A benchmark evaluating models on questions that require deep reasoning.

Systems (99)

Model Accuracy Cost In/Out Latency
Gemini 3.1 Pro Preview (02/26) 95.45%± 1.05 $2 / $12 65.76s
Gemini 3 Pro (11/25) 91.67%± 1.39 $2 / $12 61.34s
GPT 5.2 91.67%± 1.84 $1.75 / $14 173.21s
GPT 5.4 91.67%± 1.91 $2.5 / $15 202.95s
Claude Opus 4.7 89.90%± 1.56 $5 / $25 199.85s
Muse Spark 89.65%± 1.55 N/A 94.58s
Claude Opus 4.6 (Thinking) 89.65%± 1.19 $5 / $25 319.02s
Grok 4.20 (Reasoning) 88.64%± 1.59 $2 / $6 59.42s
Grok 4 88.13%± 1.63 $3 / $15 115.52s
Gemini 3 Flash (12/25) 87.88%± 1.64 $0.5 / $3 83.84s

Key Takeaways

  • In general, GPQA is largely saturated - progress on the benchmark has approached an asymptote and it seems any incremental improvements will be quite small.

Results

Input Cost Accuracy

  • AI21 Labs
  • Alibaba
  • Anthropic
  • Cohere
  • DeepSeek
  • Google
  • Kimi
  • Meta
  • MiniMax
  • Mistral
  • Moonshot AI
  • NVIDIA
  • OpenAI
  • xAI
  • Xiaomi
  • zAI

Dataset and Context

The Graduate-Level Google-Proof Q&A (GPQA) benchmark (paper) is a public academic benchmark commonly used to measure models’ general question-answering performance. It evaluates language models on challenging, graduate-level questions across STEM fields. When created, the questions were specifically designed to be “Google-proof” - they require deep understanding and reasoning rather than fact recall or search.

As in the original paper, this benchmark uses two evaluation approaches:

  • Zero-shot chain-of-thought: Models are asked to solve the problems with instruction to explain their reasoning steps.
  • Few-shot chain-of-thought: In addition to the instructions, models are provided with 5 example questions and answers.

We focus on the “diamond” subset of 198 questions - these were questions in which expert validators answered correctly and no more than one out of three non-experts answered correctly. They are both challenging but have unambiguous answers.

Additional Notes

Methodology

We only used the diamond subset (198 questions) as the evaluation set - these were the questions that were most challenging for non-experts but had the highest agreement rate between experts. We replicated the original paper’s prompting and response parsing techniques as closely as possible, although for zero-shot CoT, we only prompt the model once, rather than asking the question across two separate queries. Here are the prompts used:

Zero-shot CoT Prompt:

What is the correct answer to this question:

Two quantum states with energies E1 and E2 have a lifetime of 10^-9 sec and 10^-8 sec, respectively. We want to clearly distinguish these two energy levels. Which one of the following options could be their energy difference so that they can be clearly resolved?

Choices:
(A) 10^-8 eV
(B) 10^-9 eV
(C) 10^-4 eV
(D) 10^-11 eV

Reason through your answer step-by-step. Then, based on your reasoning, provide the single most likely answer choice. Answer in the format "The correct answer is (insert answer here)."

Few-shot CoT Prompt:

Here are some example questions from experts. An explanation is given before the final answer. Answer the final question yourself, giving your reasoning beforehand.
...

This structured content retains key information regarding GPQA, the models analyzed, the evaluation methodology, and prompts used while omitting non-semantic elements.